Structure-Preserving Reduced-Order Modeling via Low-Rank Transport Signatures

Fengyan Li; Jiajia Yu; Jingwei Hu; Shanyin Tong; Yunan Yang; Zhaiming Shen

Structure-Preserving Reduced-Order Modeling via Low-Rank Transport Signatures

T0 review · 2 major / 0 minor · reviewed 2026-07-03 · grok-4.3

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Transport signatures from Kantorovich potentials of a fixed reference density enable low-rank reduced-order models for parametrized density PDEs with explicit Wasserstein error control.

desk verdict The paper gives a transport-signature ROM that preserves mass and claims lower rank plus a separated Wasserstein error bound, but the fixed-reference assumption is the part that needs the most checking. read the letter →

arxiv 2607.01696 v1 pith:WM6R2P4R submitted 2026-07-02 math.NA cs.NA

Jiajia Yu , Jingwei Hu , Fengyan Li , Shanyin Tong , Yunan Yang , Zhaiming Shen This is my paper · ORCID

classification math.NAcs.NA

keywords reducedordermodelingoptimaltransportKantorovichpotentialsWassersteinmetricdensitysolutionsparametrizedPDElowrankapproximationneuralnetworks

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper introduces a reduced-order modeling technique for parametrized partial differential equations that produce density-valued solutions, a setting where standard linear reduced-order models often require high ranks especially under strong transport. Each density is encoded by the Kantorovich potential that pushes a single fixed reference density forward to the solution density; these potentials are then transformed into transport signatures by applying a weighted Laplacian defined with respect to the reference measure. The signatures form a matrix indexed by parameters and spatial points that admits a low-rank skeleton decomposition selected by a maximal-volume criterion, after which a neural network learns the map from parameters to the low-rank coefficients. Reconstruction always pushes the reference density forward, enforcing mass conservation, and the method supplies a mean-squared Wasserstein error bound that decomposes the total error into low-rank, discretization, sampling, and learning contributions.

What carries the argument

The transport signature, obtained by applying a weighted Laplacian associated with the reference measure to the Kantorovich potential that transports the reference density to a target density.

What would settle it

If the transport signatures extracted from the two-dimensional continuity equation example require a rank comparable to that of the original density snapshots in order to meet a prescribed Wasserstein tolerance, the advantage claimed for the signature representation would be falsified.

Watch

Extended reading notes

Core claim

The paper establishes that the map from densities to transport signatures produces a representation in which the low-rank structure is substantially better than that of the raw density fields, while the reconstruction procedure automatically preserves mass and the total mean-squared Wasserstein error can be bounded by controlling the separate contributions from rank truncation, spatial discretization, sampling of the parameter domain, and neural-network learning of the coefficient map.

Load-bearing premise

A single fixed reference density exists such that the Kantorovich potentials transporting it to each member of the solution family admit an effective low-dimensional representation after the weighted Laplacian transform.

Editorial extensions

If this is right

Low-rank approximation of the signature matrix followed by neural-network evaluation yields an efficient non-intrusive surrogate.
Push-forward reconstruction guarantees that every reconstructed density integrates to one.
The error bound isolates the effect of each approximation stage on the final Wasserstein distance.
Numerical tests on a two-dimensional continuity equation confirm that the required rank is markedly smaller than for direct density snapshots.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The framework could be combined with adaptive choice of the reference density to further reduce the observed rank in families of solutions that vary strongly.
Replacing the neural network with other regression techniques would leave the structure-preserving and error-bound properties intact.
The same signature construction might apply to other optimal-transport problems in which linear subspaces fail to capture transport-dominated behavior.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Desk Editor's Note

read the letter

The main takeaway is that this method embeds each density solution as the Kantorovich potential from one fixed reference measure, applies a weighted Laplacian to produce a signature, then extracts a low-rank skeleton with the maximal-volume criterion and fits the coefficients with a neural network. The online step pushes the reference density forward, so mass is conserved by construction. They state a mean-squared Wasserstein error bound that splits low-rank, discretization, sampling, and learning contributions, and they show on a two-dimensional continuity equation that the signatures require substantially fewer modes than the raw density snapshots.

What is new is the specific combination of the potential embedding, the Laplacian signature map, and the maxvol skeleton inside a non-intrusive ROM pipeline for transport-dominated density problems. The numerical rank reduction and the explicit error separation are the concrete pieces that stand out.

The fixed-reference assumption is the clearest soft spot. If the solution densities involve strongly varying supports or transport directions, a single reference may not produce signatures that stay low-rank, and the claimed separation in the bound could depend on that representation holding. The abstract presents the step without conditions on the reference or an independent proof that the Laplacian controls the Wasserstein distance apart from the low-rank step, so the stress-test concern lands on the given material.

This is aimed at researchers who build reduced models for hyperbolic or continuity equations where linear methods lose structure. A reader already working with optimal transport or Wasserstein distances in numerics would get the most from the pipeline and the bound.

The work is coherent enough on its own terms to go to a serious referee.

Referee Report

2 major / 0 minor

Summary. The paper introduces an optimal-transport-based reduced-order modeling framework for parametrized PDEs whose solutions are densities. Each density is represented via the Kantorovich potential that transports a single fixed reference measure to the target density; these potentials are then mapped to transport signatures in a Hilbert space by a weighted Laplacian operator associated with the reference. A low-rank skeleton is extracted from the resulting continuous matrix via the maximal-volume criterion, a neural network learns the parameter-to-coefficient map, and the solution is reconstructed by push-forward of the reference (ensuring mass preservation). The authors claim a mean-squared Wasserstein error bound that separates low-rank approximation, discretization, sampling, and learning contributions, together with a numerical demonstration on a two-dimensional continuity equation in which the transport signatures exhibit substantially lower rank than the original density snapshots.

Significance. If the fixed-reference representation is valid and the error separation holds, the approach would supply a structure-preserving, mass-conserving alternative to linear ROMs precisely in the transport-dominated regimes where the latter typically fail. The explicit separation of error sources and the use of a maxvol skeleton are concrete strengths that would make the method attractive for non-intrusive reduced modeling of conservation laws.

major comments (2)

[Abstract] Abstract (representation step): the entire low-rank claim and the separation in the Wasserstein error bound rest on the existence of a single fixed reference density μ such that the family of solution densities admits a well-behaved representation by Kantorovich potentials transporting μ. No explicit hypothesis on μ (e.g., a uniform bound on the support or on the transport cost) is stated that would guarantee this representation remains low-rank when supports or transport directions vary strongly across the parameter domain.
[Abstract] Abstract (error bound): the claimed mean-squared Wasserstein bound is asserted to separate low-rank, discretization, sampling, and learning errors, yet the text supplies no indication that the control of the Wasserstein distance by the weighted-Laplacian signature map has been shown to be independent of the subsequent maxvol low-rank step; if the two are coupled, the separation asserted in the bound does not follow.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful reading and the two major comments, which identify points where the presentation can be strengthened. We address each comment below and will revise the manuscript accordingly.

read point-by-point responses

Referee: [Abstract] Abstract (representation step): the entire low-rank claim and the separation in the Wasserstein error bound rest on the existence of a single fixed reference density μ such that the family of solution densities admits a well-behaved representation by Kantorovich potentials transporting μ. No explicit hypothesis on μ (e.g., a uniform bound on the support or on the transport cost) is stated that would guarantee this representation remains low-rank when supports or transport directions vary strongly across the parameter domain.

Authors: We agree that an explicit hypothesis on the reference measure μ is required to guarantee the low-rank property under varying supports. While the manuscript discusses the choice of μ in Section 2 and assumes a fixed reference throughout, no standing assumption is stated in the abstract. In the revision we will introduce Assumption 2.1 requiring that all solution densities have supports contained in a fixed compact set Ω and that W_2(ρ_θ, μ) is uniformly bounded for θ in the parameter domain. This ensures the Kantorovich potentials belong to a bounded set in H^1(Ω), from which low-rank structure follows by compactness. The abstract will be updated to reference this assumption. revision: yes
Referee: [Abstract] Abstract (error bound): the claimed mean-squared Wasserstein bound is asserted to separate low-rank, discretization, sampling, and learning errors, yet the text supplies no indication that the control of the Wasserstein distance by the weighted-Laplacian signature map has been shown to be independent of the subsequent maxvol low-rank step; if the two are coupled, the separation asserted in the bound does not follow.

Authors: The separation is established in the manuscript as follows: Lemma 3.4 shows that the weighted-Laplacian signature map controls the Wasserstein distance using only properties of μ and the Laplacian operator, with no dependence on any low-rank approximation. The low-rank error is then bounded separately in the signature space (Theorem 4.3) via the maxvol criterion, after which the triangle inequality yields the total mean-squared Wasserstein bound. The maxvol step is simply one concrete low-rank projector and does not couple back into the signature-to-Wasserstein control. To make this logical independence explicit, we will add a clarifying remark immediately after the statement of the error bound in the abstract and a short pointer to Lemma 3.4 in the introduction. revision: yes

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity; derivation builds on external OT theory and standard low-rank techniques

full rationale

The paper defines transport signatures from Kantorovich potentials relative to a fixed reference density followed by a weighted Laplacian, then applies maxvol skeletonization and neural-network coefficient learning. The claimed mean-squared Wasserstein error bound explicitly separates low-rank, discretization, sampling, and learning contributions, and the rank-reduction observation is presented as an empirical outcome on the continuity-equation example. None of these steps reduce by construction to a quantity defined from the output itself, nor rely on load-bearing self-citations; the fixed-reference assumption is stated as a modeling hypothesis rather than derived from the method. The overall chain therefore remains self-contained against external optimal-transport and approximation-theory results.

Assumptions & free parameters 2 free parameters · 1 assumptions · 0 invented entities

Abstract-only review; ledger populated from explicitly mentioned components only.

free parameters (2)

reference density
Fixed reference chosen once for the entire parametrized family; its selection affects all subsequent signatures.
skeleton rank
Chosen via maximal-volume criterion; directly controls the low-rank truncation error term.

assumptions (1)

domain assumption Existence of Kantorovich potentials between the reference and each solution density
Invoked in the representation step of the abstract.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Structure-Preserving Reduced-Order Modeling via Low-Rank Transport Signatures." pith.science (2026). https://pith.science/paper/WM6R2P4R

@misc{pith2026260701696,
  author       = {Pith},
  title        = {Pith review of: Structure-Preserving Reduced-Order Modeling via Low-Rank Transport Signatures},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WM6R2P4R}},
  note         = {Machine review of arXiv:2607.01696}
}

read the original abstract

Parametrized PDEs with density-valued solutions are often difficult to approximate with classical linear reduced-order models, especially in transport-dominated regimes. We introduce an optimal-transport-based reduced-order modeling that represents each density by the Kantorovich potential transporting a fixed reference density to the target density, and then maps these potentials to transport signatures using a weighted Laplacian associated with the reference measure. This embeds the density-valued solution map in a Hilbert space while preserving control of the induced transport maps and Wasserstein error. We treat the signature map as a continuous matrix indexed by parameters and space, construct a low-rank skeleton decomposition using a maximal-volume criterion, and learn the parameter-to-coefficient map with a neural network for efficient non-intrusive online evaluation. The reconstructed solution is obtained by pushing forward the reference density, so mass preservation is built into the method. We prove a mean-squared Wasserstein error bound separating low-rank approximation, discretization, sampling, and learning errors, and demonstrate the method on a two-dimensional continuity equation, where transport signatures yield substantially lower-rank structure than the original density snapshots.

Figures

Figures reproduced from arXiv: 2607.01696 by the authors.

**Figure 2.** Singular value decay pattern of probability density [PITH_FULL_IMAGE:figures/full_fig_p028_2.png] view at source ↗

**Figure 3.** Samples of training data. Top row: FOM results; bottom row: ROM results. [PITH_FULL_IMAGE:figures/full_fig_p029_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Samples of testing data. Top row: FOM results; bottom row: ROM results. [PITH_FULL_IMAGE:figures/full_fig_p029_4.png]

**Figure 5.** Figure 5: Numerical results for new parameters. Top left: training parameters and the newly [PITH_FULL_IMAGE:figures/full_fig_p030_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

32 extracted references · 32 canonical work pages

[1]

Maximal volume matrix cross ap- proximation for image compression and least squares solution

Kenneth Allen, Ming-Jun Lai, and Zhaiming Shen. Maximal volume matrix cross ap- proximation for image compression and least squares solution. Advances in Computa- tional Mathematics , 50(5):102, 2024

work page 2024
[2]

Gradient flows

Luigi Ambrosio, Nicola Gigli, and Giuseppe Savare. Gradient flows . Lectures in Math- ematics. ETH Zürich. Birkhauser Verlag AG, Basel, Switzerland, 2 edition, December 2008

work page 2008
[3]

Wasserstein model reduction approach for parametrized flow problems in porous media

Beatrice Battisti, Tobias Blickhan, Guillaume Enchéry, Virginie Ehrlacher, Damiano 30 Lombardi, and Olga Mula. Wasserstein model reduction approach for parametrized flow problems in porous media. ESAIM: Proceedings and Surveys , 73:28–47, 2023

work page 2023
[4]

Model Reduction and Approximation: Theory and Algorithms , volume 15 of Computational Science and Engineering

Peter Benner, Albert Cohen, Mario Ohlberger, and Karen Willcox, editors. Model Reduction and Approximation: Theory and Algorithms , volume 15 of Computational Science and Engineering. Society for Industrial and Applied Mathematics, Philadelphia, PA, 2017

work page 2017
[5]

Bernard, A

F. Bernard, A. Iollo, and S. Riffaud. Reduced-order model for the BGK equation based on POD and optimal transport. Journal of Computational Physics , 373:545–570, 2018

work page 2018
[6]

A registration method for reduced basis problems using linear optimal transport

Tobias Blickhan. A registration method for reduced basis problems using linear optimal transport. SIAM Journal on Scientific Computing , 46(5):A3177–A3204, 2024

work page 2024
[7]

Polar factorization and monotone rearrangement of vector-valued func- tions

Yann Brenier. Polar factorization and monotone rearrangement of vector-valued func- tions. Communications on pure and applied mathematics , 44(4):375–417, 1991

work page 1991
[8]

Nonparametric regression on low-dimensional manifolds using deep ReLU networks: Function approximation and statistical recovery

Minshuo Chen, Haoming Jiang, Wenjing Liao, and Tuo Zhao. Nonparametric regression on low-dimensional manifolds using deep ReLU networks: Function approximation and statistical recovery. Information and Inference: A Journal of the IMA , 11(4):1203–1253, 2022

work page 2022

Show all 32 references

[9]

Double greedy algorithms: Reduced basis methods for transport dominated problems

Wolfgang Dahmen, Christian Plesken, and Gerrit Welper. Double greedy algorithms: Reduced basis methods for transport dominated problems. ESAIM: Mathematical Mod- elling and Numerical Analysis , 48(3):623–663, 2014

2014
[10]

Sparse wasserstein barycenters and appli- cation to reduced order modeling

Minh-Hieu Do, Jean Feydy, and Olga Mula. Sparse wasserstein barycenters and appli- cation to reduced order modeling. Journal of Scientific Computing , 102:64, 2025

2025
[11]

Non- linear model reduction on metric spaces

Virginie Ehrlacher, Damiano Lombardi, Olga Mula, and François-Xavier Vialard. Non- linear model reduction on metric spaces. application to one-dimensional conservative PDEs in Wasserstein spaces. ESAIM: Mathematical Modelling and Numerical Analysis , 54(6):2159–2197, 2020

2020
[12]

A comprehensive deep learning-based approach to reduced order modeling of nonlinear time-dependent parametrized PDEs

Stefania Fresca, Luca Dedè, and Andrea Manzoni. A comprehensive deep learning-based approach to reduced order modeling of nonlinear time-dependent parametrized PDEs. Journal of Scientific Computing , 87(2):61, 2021

2021
[13]

Goreinov, Ivan V

Sergei A. Goreinov, Ivan V. Oseledets, Dimitry V. Savostyanov, Eugene E. Tyrtysh- nikov, and Nikolay L. Zamarashkin. How to find a good submatrix. In Matrix Methods: Theory, Algorithms and Applications: Dedicated to the Memory of Gene Golub , pages 247–256. World Scientific, 2010

2010
[14]

Goreinov and Eugene E

Sergei A. Goreinov and Eugene E. Tyrtyshnikov. The maximal-volume concept in approximation by low-rank matrices. Contemporary Mathematics, 268:47–51, 2001

2001
[15]

Goreinov, Eugene E

Sergei A. Goreinov, Eugene E. Tyrtyshnikov, and Nickolai L. Zamarashkin. A theory of pseudoskeleton approximations. Linear Algebra and its Applications , 261(1-3):1–21, 1997. 31

1997
[16]

Hesthaven, Cecilia Pagliantini, and Gianluigi Rozza

Jan S. Hesthaven, Cecilia Pagliantini, and Gianluigi Rozza. Reduced basis methods for time-dependent problems. Acta Numerica, 31:265–345, 2022

2022
[17]

Nonlinear model reduc- tion for transport-dominated problems

Jan S Hesthaven, Benjamin Peherstorfer, and Benjamin Unger. Nonlinear model reduc- tion for transport-dominated problems. arXiv preprint arXiv:2602.01397 , 2026

2026
[18]

Hesthaven, Gianluigi Rozza, and Benjamin Stamm

Jan S. Hesthaven, Gianluigi Rozza, and Benjamin Stamm. Certified Reduced Basis Methods for Parametrized Partial Differential Equations. SpringerBriefs in Mathematics. Springer, Cham, 2016

2016
[19]

A fast approach to optimal transport: the back-and- forth method

Matt Jacobs and Flavien Léger. A fast approach to optimal transport: the back-and- forth method. Numerische Mathematik , 146(3):513–544, 2020

2020
[20]

The variational formulation of the Fokker–Planck equation

Richard Jordan, David Kinderlehrer, and Felix Otto. The variational formulation of the Fokker–Planck equation. SIAM Journal on Mathematical Analysis , 29(1):1–17, 1998

1998
[21]

Optimal transport-based displacement interpolation with data augmentation for reduced order modeling of nonlinear dynamical systems

Moaad Khamlich, Federico Pichi, Michele Girfoglio, Annalisa Quaini, and Gianluigi Rozza. Optimal transport-based displacement interpolation with data augmentation for reduced order modeling of nonlinear dynamical systems. Journal of Computational Physics, 531:113938, 2025

2025
[22]

Soheil Kolouri, Se Rim Park, Matthew Thorpe, Dejan Slepčev, and Gustavo K. Rohde. Optimal mass transport: Signal processing and machine-learning applications. IEEE Signal Processing Magazine , 34(4):43–59, 2017

2017
[23]

Tosun, John A

Soheil Kolouri, Akif B. Tosun, John A. Ozolek, and Gustavo K. Rohde. A continu- ous linear optimal transport approach for pattern analysis in image datasets. Pattern Recognition, 51:453–462, 2016

2016
[24]

Quantitative stability of optimal transport maps and linearization of the 2-wasserstein space

Quentin Mérigot, Alex Delalande, and Frédéric Chazal. Quantitative stability of optimal transport maps and linearization of the 2-wasserstein space. In Proceedings of the Twenty Third International Conference on Artificial Intelligence and Statistics , volume 108 of Proceeding...

2020
[25]

A reduced basis method by means of transport maps for a fluid–structure interaction problem with slowly decaying kolmogorov n-width

Monica Nonino, Francesco Ballarin, Gianluigi Rozza, and Yvon Maday. A reduced basis method by means of transport maps for a fluid–structure interaction problem with slowly decaying kolmogorov n-width. Advances in Computational Science and Engineering, 1(1):36–58, 2023

2023
[26]

Model reduction for transport-dominated problems via on- line adaptive bases and adaptive sampling

Benjamin Peherstorfer. Model reduction for transport-dominated problems via on- line adaptive bases and adaptive sampling. SIAM Journal on Scientific Computing , 42(5):A2803–A2836, 2020

2020
[27]

Computational optimal transport

Gabriel Peyré and Marco Cuturi. Computational optimal transport. Foundations and Trends in Machine Learning , 11(5–6):355–607, 2019

2019
[28]

Donsub Rim, Benjamin Peherstorfer, and Kyle T. Mandli. Manifold approximations via transported subspaces: Model reduction for transport-dominated problems. SIAM Journal on Scientific Computing , 45(1):A170–A199, 2023. 32

2023
[29]

Optimal Transport for Applied Mathematicians: Calculus of Variations, PDEs, and Modeling , volume 87 of Progress in Nonlinear Differential Equa- tions and Their Applications

Filippo Santambrogio. Optimal Transport for Applied Mathematicians: Calculus of Variations, PDEs, and Modeling , volume 87 of Progress in Nonlinear Differential Equa- tions and Their Applications . Birkhäuser, Cham, 2015

2015
[30]

Trefethen

Alex Townsend and Lloyd N. Trefethen. Continuous analogues of matrix factorizations. Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences , 471(2173):20140585, 2015

2015
[31]

Topics in Optimal Transportation , volume 58 of Graduate Studies in Mathematics

Cédric Villani. Topics in Optimal Transportation , volume 58 of Graduate Studies in Mathematics. American Mathematical Society, Providence, RI, 2003

2003
[32]

Ozolek, and Gustavo K

Wei Wang, Dejan Slepčev, Saurav Basu, John A. Ozolek, and Gustavo K. Rohde. A linear optimal transportation framework for quantifying and visualizing variations in sets of images. International Journal of Computer Vision , 101(2):254–269, 2013. 33

2013

Pith tools

Reviewed July 3, 2026 · model on record in the stance chip above.

[1] [1]

Maximal volume matrix cross ap- proximation for image compression and least squares solution

Kenneth Allen, Ming-Jun Lai, and Zhaiming Shen. Maximal volume matrix cross ap- proximation for image compression and least squares solution. Advances in Computa- tional Mathematics , 50(5):102, 2024

work page 2024

[2] [2]

Gradient flows

Luigi Ambrosio, Nicola Gigli, and Giuseppe Savare. Gradient flows . Lectures in Math- ematics. ETH Zürich. Birkhauser Verlag AG, Basel, Switzerland, 2 edition, December 2008

work page 2008

[3] [3]

Wasserstein model reduction approach for parametrized flow problems in porous media

Beatrice Battisti, Tobias Blickhan, Guillaume Enchéry, Virginie Ehrlacher, Damiano 30 Lombardi, and Olga Mula. Wasserstein model reduction approach for parametrized flow problems in porous media. ESAIM: Proceedings and Surveys , 73:28–47, 2023

work page 2023

[4] [4]

Model Reduction and Approximation: Theory and Algorithms , volume 15 of Computational Science and Engineering

Peter Benner, Albert Cohen, Mario Ohlberger, and Karen Willcox, editors. Model Reduction and Approximation: Theory and Algorithms , volume 15 of Computational Science and Engineering. Society for Industrial and Applied Mathematics, Philadelphia, PA, 2017

work page 2017

[5] [5]

Bernard, A

F. Bernard, A. Iollo, and S. Riffaud. Reduced-order model for the BGK equation based on POD and optimal transport. Journal of Computational Physics , 373:545–570, 2018

work page 2018

[6] [6]

A registration method for reduced basis problems using linear optimal transport

Tobias Blickhan. A registration method for reduced basis problems using linear optimal transport. SIAM Journal on Scientific Computing , 46(5):A3177–A3204, 2024

work page 2024

[7] [7]

Polar factorization and monotone rearrangement of vector-valued func- tions

Yann Brenier. Polar factorization and monotone rearrangement of vector-valued func- tions. Communications on pure and applied mathematics , 44(4):375–417, 1991

work page 1991

[8] [8]

Nonparametric regression on low-dimensional manifolds using deep ReLU networks: Function approximation and statistical recovery

Minshuo Chen, Haoming Jiang, Wenjing Liao, and Tuo Zhao. Nonparametric regression on low-dimensional manifolds using deep ReLU networks: Function approximation and statistical recovery. Information and Inference: A Journal of the IMA , 11(4):1203–1253, 2022

work page 2022